Statistical Tests for Cross-Validation of Kriging Models

Kriging or Gaussian process models are popular metamodels (surrogate emulators) of simulation models; these give predictors for input combinations that not simulated. To validate computationally expensive models, the analysts often apply efficient cross-validation. In this paper, we derive new statistical tests so-called leave-one-out Graphically, present as scatterplots augmented with confidence intervals use estimated variances predictors. estimate true predictors, might bootstrapping. Like other tests, our tests—with without bootstrapping—have type I and II error probabilities; to probabilities, Monte Carlo experiments. We also such experiments investigate convergence. illustrate application (i) an example two inputs (ii) borehole eight inputs. Summary Contribution: Simulation very in operations research (OR) known computer simulations A topic is design analysis This paper focuses on methods cross-validation applied OR. More specifically, provides following; (1) basic variant a test leave-one–out cross-validation; (2) bootstrap method estimation variance predictor; (3) evaluation consistency predictor, convergence Studentized prediction standard normal variable, expected experimentwise rate prespecified nominal value. The illustrated through examples, including model.